Architecture cost review
Cloudflare Clef Pricing: What Each Clef Tier Costs per Decision
Clef has three tiers on OpenRouter and Apache 2.0 weights on Hugging Face. The hosted price spread is five-fold, and the cheapest tier has the smallest window.
Direct answer
Buy Clef Flash: $0.038 per million input tokens on OpenRouter, $19 per million 500-token decisions, a little under Jev's $21. Standard Clef costs $0.192 per million input tokens, or $96 per million decisions, and earns that only when the state includes images. Clef Omni, at $0.15, is oddly cheaper than Clef. Decider Flash still undercuts every Clef tier at $5. The open weights are the real draw.
Decision summary
| Decision area | What matters |
|---|---|
| Clef Flash | $0.038/M input · $0 output · 24,576 tokens · 9B on Qwen3.5-9B |
| Clef | $0.192/M · $0 · 65,537 tokens · 27B multimodal on Qwen3.8-27B |
| Clef Omni | $0.15/M · $0 · 65,536 tokens · Qwen3-Omni-30B-A3B, 3B active |
| Per 1M decisions at 300 / 500 / 2,000 tokens | Flash $11.40 / $19 / $76 · Clef $57.60 / $96 / $384 · Omni $45 / $75 / $300 |
| Same volumes on Jev | $12.60 / $21 / $84 |
| Same volumes on Decider Flash | $3 / $5 / $20 |
| Weights | Clef and Clef Flash: Apache 2.0 on Hugging Face (Developers Digest, Classmethod) |
Three tiers, one sensible default
Cloudflare released Clef and Clef Flash on 1 October 2026, and OpenRouter added Clef Omni on 9 October. All three answer typed questions with probabilities and charge nothing for output. Each sits on a Qwen base: Clef on Qwen3.8-27B, Clef Flash on Qwen3.5-9B and Clef Omni on Qwen3-Omni-30B-A3B.
Clef Flash is the one to start on. It is the only tier priced below Jev, and it handles the short text states most decision points use.
Per-decision cost at three state sizes
Multiply state tokens by the rate. At 300 tokens: Clef Flash 300 × $0.038 = $11.40 per million decisions, Clef Omni $45, Clef $57.60. At 500 tokens: $19, $75 and $96. At 2,000 tokens: $76, $300 and $384.
The window caps what Flash can do. Its 24,576 tokens hold a long email thread, which at 20,000 tokens costs $0.76 per thousand decisions. A 50,000-token state needs Clef or Omni, at $9.60 or $7.50 per thousand.
Why Clef Omni costs less than Clef
OpenRouter lists Clef Omni at $0.15 and Clef at $0.192, so the newer, larger-sounding model is 22% cheaper. Omni is a mixture-of-experts model with 30B total and 3B active parameters, which may explain a lower serving cost. Cloudflare has not said, and we found no test of Omni's accuracy.
Classmethod confirms Clef accepts up to four images per request. The Omni listing does not spell out its inputs beyond the base model's name. Test Omni on your own image cases before moving Clef traffic to it to save $21 per million decisions.
Open weights and running Clef yourself
Developers Digest and Classmethod both report Clef and Clef Flash weights on Hugging Face under Apache 2.0. Classmethod notes you can run them on your own GPU. Developers Digest adds that the training data and pipeline are not published.
We have no self-hosting cost to quote, and we will not guess one. Price it the way our Laya page does: monthly GPU cost ÷ hosted cost per decision. Against Clef Flash at $19 per million 500-token decisions, a $252-a-month GPU breaks even at about 13.3 million decisions a month ($252 ÷ $19).
Cloudflare also serves Clef on Workers AI at its own rates. Early-October write-ups quoted rates that differ from OpenRouter's 3 November listing. Check Cloudflare's model page before assuming either is cheaper.
Clef against the cheaper field
At 10 million 500-token decisions a month, Clef Flash costs $190, Jev $210 and Decider Flash $50. Decider Flash also reads 262,144 tokens against Flash's 24,576. On price and window alone, Clef Flash loses to Decider Flash.
Clef's case is the weights and the evidence. It is one of the few decision-model families with Apache 2.0 weights and an independent test. In a 117-case DEV Community test, Clef scored 92.3% and Clef Flash 82.1% at Jev's thresholds. With each model's own threshold, Flash reached 90.6%. That test also found Clef returned the same probability, to four decimals, on all 351 calls.
Monthly budgets by tier
At a 500-token state, multiply monthly decisions by 500 tokens and by the rate. Clef Flash costs $1.90 at 100,000 decisions, $19 at 1 million, $190 at 10 million and $950 at 50 million. Standard Clef costs $9.60, $96, $960 and $4,800 at the same volumes. Clef Omni costs $7.50, $75, $750 and $3,750.
OpenRouter's 5.5% credit fee adds a little: Clef Flash at 10 million becomes $200.45 ($190 × 1.055). Standard Clef at 50 million becomes $5,064. At that level the hosted bill is large enough to justify pricing your own GPUs, which the open weights allow.
Moving from Jev to Clef Flash
Classmethod reports Clef speaks Jev's API: change the endpoint, key and model name. Through OpenRouter, only the model name changes. If you call Workers AI's REST API directly, results come wrapped in a result field.
The window is the hard limit. Jev reads 262,144 tokens and Clef Flash 24,576. Before moving, log the size of every state sent to Jev for a week. Anything over 24,576 tokens must be trimmed, go to standard Clef, or stay on Jev. Then re-pick thresholds: in the DEV Community test, Clef Flash scored 82.1% on Jev's thresholds and 90.6% on its own.
What to check before you commit to Clef
Each check below comes from the listings and write-ups cited on this page.
- Your largest state against 24,576 tokens (Flash) or 65,537 (Clef)
- Whether your images meet Clef's limits: up to four per request, PNG, JPEG or WebP, 4 MiB each, no remote URLs (Classmethod)
- Cloudflare's current Workers AI rate against OpenRouter's, if you plan to call Cloudflare directly
- Your own latency from your region; Cloudflare's figures are vendor-run
Key takeaways
- →Buy Clef Flash: $0.038 per million input tokens on OpenRouter, $19 per million 500-token decisions, a little under Jev's $21. Standard Clef costs $0.192 per million input tokens, or $96 per million decisions, and earns that only when the state includes images. Clef Omni, at $0.15, is oddly cheaper than Clef. Decider Flash still undercuts every Clef tier at $5. The open weights are the real draw.
- →Start on Clef Flash for text states under 24,000 tokens, move only image-bearing decisions to Clef, and price self-hosting once volume is steady.
- →Cloudflare's own Workers AI rates, quoted in early-October write-ups, differ from OpenRouter's; we list only OpenRouter's 3 November prices.
What users report
Users find full Clef expensive next to Jev, while Clef Flash's lower price and self-hosting make it attractive for cheap, always-on checks.
Paraphrased from public posts, with the date and a link to each original. These are other people's accounts, not our tests, and we have not verified each one.
One user noted Clef costs $0.24 per million input tokens, about 6x Jev, while Clef Flash at $0.09 was "way more competitive".
Billing · Clef · Hacker News, Oct 2026
A commenter worked out that a million 300-token decisions would cost about $72 on Clef versus about $12.60 on Jev.
Billing · Clef · Hacker News, Oct 2026
A developer runs screenshots and app state through Clef with questionnaires in CI, saying it "costs pennies or less" and was far more cost effective than Claude or GPT models.
Value · Clef · Hacker News, Nov 2026
One IoT developer runs Clef Flash first in an always-on local stack, saying small models are fast and light on memory, for work they'd never pay frontier prices for.
Value · Clef Flash · Hacker News, Oct 2026
One user said competitors were not actually cheaper at scale, reporting that both OpenAI and Cloudflare came in at twice Jev's cost.
Billing · Clef · Hacker News, Oct 2026
How this page was prepared
Prices on this page come from OpenRouter's live models API, read on 3 November 2026. Cost per decision is state tokens multiplied by the input rate, because output is free on every decision model listed; small-LLM costs add the stated label length at that model's output rate. OpenRouter's credit fee is shown separately where it matters. We ran no private benchmark for these pages: accuracy and latency figures are credited to the vendor or third party that measured them, and vendor figures are labelled as vendor claims.
- OpenRouter live models API (checked 3 Nov 2026) ↗
- OpenRouter — cloudflare/clef model page ↗
- OpenRouter — cloudflare/clef-flash model page ↗
- Developers Digest — Cloudflare Clef decision models (2026) ↗
- DevelopersIO (Classmethod) — Cloudflare Clef overview ↗
- DEV Community (ankushkun) — Clef beat Jev on Jev's home turf ↗
Frequently asked questions
How much does Cloudflare Clef cost?
On OpenRouter on 3 November 2026: Clef $0.192, Clef Omni $0.15 and Clef Flash $0.038 per million input tokens, with free output.
Which Clef should I use?
Clef Flash for text states under about 24,000 tokens. Clef when the state includes images. Test Omni before trusting it with Clef's work.
Is Clef open source?
Clef and Clef Flash weights are Apache 2.0 on Hugging Face, per Developers Digest and Classmethod. The training data and pipeline are not published.
Is Clef Flash cheaper than Jev?
Per token, yes: $0.038 against $0.042, or $19 against $21 per million 500-token decisions. Jev's window is more than ten times larger.
What does Clef cost on Workers AI?
Cloudflare sets its own Workers AI rates, and early-October reports differ from OpenRouter's November listing. Check Cloudflare's model page for the current figure.
Is Clef Flash good enough to replace standard Clef?
For text-only states it may be. In the DEV Community test, Flash scored 82.1% against Clef's 92.3% at Jev's thresholds, and 90.6% against 93.2% with each model's own threshold. At $19 against $96 per million decisions, test Flash on your cases first.
Does Clef charge for output?
No. All three tiers list $0 output on OpenRouter on 3 November 2026, and Developers Digest reports Workers AI lists no output charge either. The bill is state tokens only.
Can I fine-tune Clef on my own labels?
The Apache 2.0 licence allows you to modify the published Clef and Clef Flash weights, so you can fine-tune and run your own copy. Our sources describe no hosted fine-tuning option, and the training data is not published.